Large language models — ChatGPT, Claude, Gemini, and their successors — have permanently changed the landscape of academic writing. Understanding how to use them ethically and effectively is now a core academic skill. This guide draws on the most current research to give students a clear, practical framework.
The research context
The academic literature on AI in education has grown rapidly since 2023. Several key findings shape the guidance in this article:
Kasneci et al. (2023) reviewed the opportunities and challenges of large language models in education in Learning and Individual Differences and concluded that LLMs offer genuine pedagogical value when used to support learning processes — generating explanations, providing feedback, and enabling exploration of ideas — but create risks when they replace rather than scaffold cognitive engagement.
Cotton, Cotton & Shipway (2024) examined academic integrity in the era of ChatGPT in Innovations in Education and Teaching International and found that institutional policies were lagging behind student practice, with significant variation in what was permitted across institutions. They also found that AI detection tools were unreliable, making policy enforcement through detection largely ineffective.
Yan et al. (2024) conducted a systematic scoping review in the British Journal of Educational Technology and identified that the most significant learning cost of AI use in education was not citation plagiarism but cognitive offloading — the transfer of thinking to the AI rather than the student, which undermines the skill development that academic writing is designed to produce.
Lodge et al. (2023) in Educational Philosophy and Theory argued that assessment in higher education needs to be reconceptualised to address AI not as a threat to be policed but as a changed context that requires new approaches to evaluating learning.
The core distinction: AI as thinking partner vs. AI as writer
The most useful framework for ethical AI use in academic writing is the distinction between AI that helps you think better and AI that thinks for you.
AI as thinking partner (generally legitimate):
- Asking AI to explain a concept before you read primary sources
- Using AI to brainstorm objections to your thesis
- Asking AI for structural feedback on a draft you wrote
- Using AI to check whether your argument is clear
- Using AI-powered literature search tools (Elicit, Semantic Scholar) to surface relevant papers
AI as writer (generally problematic):
- Having AI generate essay sections you submit as your own work
- Using AI to paraphrase sources to disguise the source
- Having AI summarise papers you haven't read and citing them as if you had
- Using AI to generate a thesis statement, body paragraphs, or conclusion
The learning cost of the second category is significant. The cognitive work involved in constructing an argument — deciding what evidence supports which claim, figuring out why the evidence is relevant, working through counterarguments — is precisely what develops the analytical thinking skills that academic writing is designed to build. Offloading that work to AI produces a submitted essay without producing a better thinker. This is the deepest problem with AI-generated writing, independent of detection risk.
What universities actually permit
University AI policies vary significantly and are changing rapidly. As of 2025, most major UK and US universities have moved from blanket bans to conditional permission frameworks:
Common permitted uses (check your institution's specific policy):
- Using AI for brainstorming and idea generation (with disclosure)
- Using AI to get feedback on a draft you wrote yourself
- Using AI for literature searching (Elicit, Consensus, Semantic Scholar)
- Using AI to check grammar and clarity
- Using AI to generate code in computing courses (with limits)
Common restricted or prohibited uses:
- Using AI to generate essay content that is submitted as your own without disclosure
- Using AI in closed-book examinations
- Using AI to produce complete or near-complete drafts
- Using AI without disclosure in assessed work where the institution's policy requires it
The disclosure question: Many institutions now require students to disclose AI use in assessed work, specifying how it was used. Failing to disclose when required is itself an academic integrity violation, even if the use itself would have been permitted.
Always check the specific policy for each module and assessment. Policies vary within institutions, not just between them.
Practical ethical uses: worked examples
For literature searching
Elicit (elicit.org) and Semantic Scholar (semanticscholar.org) use AI to surface relevant papers, generate TLDR summaries of abstracts, and suggest related work. These are legitimate and increasingly expected tools in academic research. They accelerate literature searching without generating content — you still read the papers and form your own synthesis.
Use AI chatbots for broader topic mapping ("What are the main theoretical debates in X field?"), but verify everything against actual sources. AI chatbots do not have real-time access to academic databases and regularly hallucinate both citations and content.
For argument feedback
After writing a draft: "Here is my thesis and three main sub-claims. Are there significant counterarguments I haven't addressed?" or "Does this argument have obvious logical gaps?" This uses AI as a critical reader, not as a writer. You retain the intellectual work; the AI provides the perspective of a sceptical reader.
For clarity checking
"Is this paragraph clear to someone unfamiliar with the topic? What's confusing?" This is using AI as a clarity reader rather than as a writer. You keep all the ideas; the AI helps identify where the communication fails.
For referencing
Do not use AI chatbots to generate references — they hallucinate. Use the Citation Reference Formatter instead, which generates correctly formatted Harvard, APA, MLA, Chicago, and Vancouver references from source details you supply.
How to reference AI use in academic work
If your institution requires disclosure and permits some AI use, most styles are developing guidance on how to cite AI:
APA 7th edition: OpenAI. (2025). ChatGPT [Large language model]. https://chat.openai.com (The APA treats AI tools as software with no author other than the organisation.)
Harvard: OpenAI (2025) ChatGPT [Artificial intelligence]. Available at: https://chat.openai.com (Accessed: 1 June 2025).
MLA 9th edition: "Text of prompt" prompt. ChatGPT, version, OpenAI, date, URL.
Check your institution's guidance, as conventions are still evolving and some institutions have their own preferred format.
The learning argument against AI-generated writing
The most important reason to do your own academic writing is not policy compliance — it is skill development. Kasneci et al. (2023) and Yan et al. (2024) both emphasise that the cognitive struggle of constructing an argument, finding the right evidence, and working through the reasoning is itself the educational product. Students who outsource this work to AI produce the appearance of learning without the substance.
The research on deliberate practice (Ericsson et al., 1993) is relevant here: skill development requires engagement with the difficulty, not its elimination. Academic writing is hard because constructing well-evidenced analytical arguments is genuinely difficult. It is supposed to be difficult. The struggle is the learning.
For comprehensive guidance on academic writing skills, take the free Academic Writing Fundamentals course or use the Essay Structure Planner to build argument outlines before writing.
Plan your essay before you write a single word
Use the free Essay Structure Planner to build your argument outline, map PEEL paragraphs, and structure your introduction and conclusion — then take the free Academic Writing Fundamentals course for the complete essay-writing system.